> Is there an alternative they could compete on instead of the core tech?

To add to what tim333 said: The core tech doesn't exist yet, and it will be very expensive to finish developing, so the companies must be incentivized to pursue it, which pretty much necessitates competition. And indeed, the history of science and tech development shows that competition is exceedingly beneficial, and in the vast majority of cases far outweighs the minor costs of cooperability layers. The advent of a standard protocol generally marks when technological progress has stopped (although this can beneficially allow tech progress on tools built on top of the protocol).

Trouble is no one has really cracked self driving yet so it's probably premature to focus on one flawed system. Even Waymo it seems can't merge with heavy traffic https://news.ycombinator.com/item?id=19011890. We may need better AI beyond current machine learning algorithms.

Right. If the solution was just to put more miles on their vehicles. We'd see them with 10x the number of vehicles out on the road. They're not just a few months of "learning" away from cracking full self driving.

The financial incentive of getting there first is too great for them to not just throw money and hardware at the problem if that's all it took. That we see they're not doing that tells us they know they're missing some critical component. Which means all the cars out on the road now are mostly there for PR. Waymo is trying to get people comfortable with the change and to change regulations in states across the country, but they can't be "learning" anything (or they'd just scale that up and launch full self driving sooner).

No real breakthrough is needed. At this point, the current machine learning architectures and models we have are enough. The pains of R&D have been dealt with, and now it’s all about implementation and scaling, at least for the companies that have reached this point. I wouldn’t be surprised if these German car makers are still heavily behind the state of the art.

I don't think this is the case - both the breakthrough and your statement about where the car manufacturers are. First, current machine learning is flawed. I provides reasonably accurate predictions for simple to medium complex problems with manageable amount of input variables, but requires a lot of data already at the high end of that. For high confidence and high dimensionality it's already uneconomical to even get the data in amount and quality needed. These models fail ultimately with complex scenarios like day to day driving.

As you can see with Waymo and Uber even the simple grid like traffic systems in the US with wide lanes and lots of space to cross are already a problem. Here we haven't even looked at organically grown traffic systems like in Europe or the sheer craziness they call traffic in countries like India. Effectively none of these are self-driving cars as they all come with a driver who regularly has to intervene as the cars prefer to drive around the block rather than doing a left-turn. These are problems they will not be able to solve even with 100x the mileage and cars they have today - scalability is not the problem here. Or rather it's uneconomical to solve it like that - by the time you gathered the necessary amount of training data the traffic systems will likely have changed significantly.

The underlying problem is that machine learning is impossible to debug and deep learning is based on the understanding of the human brain from the early 60s. These fields has advanced significantly since then and we will need both, new algorithms and new hardware, of what will eventually become sophisticated neuromorphic networks that truly mimic the human brain capacity. As it stands today, a human needs only a couple of examples to train, a machine needs millions. This is not only a compute problem but also a sensor problem. Today's sensor don't deliver the data at the necessary fidelity fast enough. We need algorithms that can learn more effectively, not harder.

I wouldn't be surprised if the German car makers are basically where everybody else is, but just waiting until the systems gets to a level where it's economical, truly safe and well accepted to get into mass production. Until then, they are probably fine with customers of Uber / Waymo play guinea pig.

There is a huge level of information assymetry that currently exists in the self driving industry. It is very much not an equal playing field. You have to understand that certain companies are so far ahead of other ones that they are pretty much competing in distinctly different leagues at this point.

You’re perspective on what the cutting edge of machine learning is shows your understanding of the current technological landscape is lacking. It’s understandable as the developments are happening incredibly quickly and the majority of the research is done privately, and unless you follow the release of academic papers very closely, it can be very easy to fall behind. Nevertheless it’s all happening regardless of whether you’re aware of it or not.

There are many required services: testing/certification, charging/refueling, parking, mechanical maintenance, inspections, cleaning (assuming ride sharing will be the top use case), dispatch, redistribution, picking up stranded vehicles, etc. On the commercial vehicle side, things get more interesting with logistics (loading/unloading of cargo, dock operations, etc.) and to-the-customer or to-the-door delivery.

It would seem a good idea for self driving cars to talk to each other to coordinate movement. Eg four way stop signs. So they’re all not just trying to go and seeing what happens.

Also one car part suppliers (Bosch) are making components interoptablity become easier.

This would be huge, and Cadillac has already deployed this in some of their models. The basic idea is simple: your car broadcasts its position and velocity vector multiple times per second (and your "car id" is periodically randomized for security), and the signal can be detected by any receiver within 500-1000 meters.

The simplest receiver could then trigger a flashing red "collision imminent" light if things look dicey. A complex receiver could employ active controls to "platoon" a set of vehicles, co-ordinate lane changes, or negotiate intersections.

The missing links, I think, are the need for a) 1-2 meter position accuracy for platooning, and b) 100-200 cm accuracy for lane-keeping/lane changing.

You'd still need to be an active, engaged driver: more "auto-pilot" than autonomous.

Nice question. Even I would like to know more about those services

Carmakers will probably adopt the ZF system, as they did it with gearboxes. ZF's products are known to be expensive but are extremely reliable in general.